The data center industry is facing growing scrutiny over energy consumption. This isn’t surprising when you consider Gartner projects that U.S. data center electricity usage will double by 2030, driven largely by AI. For further context, the Uptime Institute reports that for every unit of energy used by IT equipment, an additional 0.56 units go toward cooling, power distribution, lighting, etc.
Meanwhile, news headlines, town halls, and online discussions about the rise in data center construction are in response to communities and utilities asking hard questions about the impact on the community, infrastructure, grid, and water supply. The tension further rises when conversations turn to halting construction of new data centers.
The questions are fair, but there are parallel conversations worth having. If we can make existing data centers more energy efficient, will that pave the way to more energy efficient data center construction without dividing cities and towns? If so, how can we do this?
Current data centers, while optimized for efficiency, can be improved. It starts with taking a closer look at the infrastructure and how it’s currently optimized.
We know that HVAC systems account for 25-40% of a data center's total energy spend. We also know that these systems are often managed the same way they have been for decades. They typically run on fixed schedules and static set points with systems designed around worst-case load assumptions. This worked well in the past when computing workloads were more predictable and less demanding than today.
Of course, upgrading existing data centers has been underway for some time. However, when upgrading a data center from CPUs to GPUs to handle the increased load, you must have certainty that the existing cooling system can support the demand.
Rethinking Data Center Cooling
Inside the data center, AI generates heat loads that are denser and less stable than traditional computing workloads. These workloads shift by rack and zone in ways that can be hard to anticipate, making it challenging to determine which zones are generating the most heat from those sitting idle.
In response, the HVAC systems are programmed to constantly cool everything at the level required by its highest-demand assumptions. This leads to chronic overcooling and significant wasted energy.
We know that power usage effectiveness (PUE) has been a meaningful metric for years. The challenge now is that the infrastructure exists to do better, but it needs better data to act on.
This partially explains why thermal based sensors have emerged as a way to better understand the impact of GPUs. Originating from a need in buildings to better understand how workplaces are being used by employees from the perspectives of productivity, collaboration, and energy efficiency, thermal sensors are providing deeper insight into the impact of inefficient cooling in data centers.
For example, many large employers learned through the use of thermal technology that parts of their offices were unoccupied for hours of the day and yet the HVAC continued to cool the areas.
The data center version of this problem carries higher stakes, because the energy costs are significantly larger and the consequences of an unexpected hot spot are more severe. Since thermal technology is inherently incapable of distinguishing individuals, this makes it easier to introduce it into the facilities management workflow.
Additionally, the anonymity of thermal technology ensures real insight based on authentic human behavior. Conversely, when workers are under a watchful eye, their actions become performative, undermining the credibility of the data.
Still, the data gap in both the workplace and data center is the same. A lack of real-time, zone-level data on what is happening inside the building as it’s happening. Without this information, facilities managers make decisions based on assumptions rather than real world conditions.
The Data Center of the Future
Thermal sensors fill that gap by detecting surface temperature and translating it into location, occupancy and activity data at the zone level. On its own, a sensor does not tell you much. However, if you connect enough of them across a facility, they become part of the building’s “nervous system,” or DNA.
This provides a real-time picture of thermal conditions that were previously unavailable. Sensor data makes it possible to be more responsive to immediate needs and spot longer-term trends before they lead to outages. This can improve energy efficiency by 20-30% by matching cooling output to actual demand rather than assumed demand.
A facility that reduces HVAC energy use by 30% achieves a meaningful reduction in demand without having to take a single server offline. For facilities managers, building engineers, and data center stakeholders, being able to address today’s data center energy efficiency needs is essential. As energy demand from AI workloads grows, optimizing energy efficiency is critical to existing and future data center planning.
